Panasonic 18650PF Li-ion Battery Data and Example FNN and LSTM Neural Network SOC Estimator Training Script

The included example script was created by Dr. Carlos Vidal and Dr. Phillip Kollmeyer at McMaster University in Hamilton, Ontario, Canada. If this script and the included dataset is utilized for any purpose, the following paper should be referenced along with this Mendeley dataset:

Carlos Vidal, Pawel Malysz, Mina Naguib, Ali Emadi, Phillip J. Kollmeyer, “Estimating battery state of charge using recurrent and non-recurrent neural networks,” Journal of Energy Storage, 2021 (see https://www.sciencedirect.com/ for complete citation information).

The example script is configured to train two different types of machine learning state of charge estimation algorithms - a feedforward neural network with filtered input values and a long short term memory (LSTM) recurrent neural network. These algorithms are described in detail in the above reference. The script trains the SOC estimator for normalized data for a Panasonic 18650PF battery dataset which can be found here: https://data.mendeley.com/datasets/wykht8y7tg/1

%*****************************************************************************************%
-Instructions for Downloading and Running the Script:
1-Select download all files from the Mendeley Data page
2-The files will be downloaded as a zip file. Unzip the file to a folder, do not modify the folder structure.
3-Open and run "LSTMvsFNN_Script_Nov_2021_V2.mlx"
4-Further instructions are included in the comments and text in the script (for better experience use the "Matlab Live Code File" with extension *.mlx).

%*****************************************************************************************%
Important - Additional notes regarding the script:
When changing the type of neural network go to line 101 and follow the instructions below.

To select Neural Network Type (line 101):

Select "1" on the drop down for LSTM
Select "2" on the drop down for FNN

Also go to line 237 and make the same selection of Neural Network Type.

%*****************************************************************************************%
Description of included files:

-All data is normalized, see "Normalization" folder for example showing how to denormalize data

-Data for -20, -10, 0, 10, and 25degC is included (in the paper, the -20degC data is not used)

-The "X" data has 7 rows, where the data in each row is as follows: { V, I, T, V_0.5mHz, I_0.5mHz, V_5mHz, I_5mHz}, where V is voltage, I is current, T is temperature, and the _0.5mHz and _5mHz data is filtered with a 1st order low pass Butterworth filter

-The "Y" data is state of charge calculated via coulomb counting

-The data is split into Train, Test, and Validation and is saved in the respective folders (see the Journal of Energy Storage paper for a description of how the data is split). Mix 1 to 4 and US06 are used for training, LA92 -10degC for validation, and LA92, NN, and UDDS for testing

-The training data is split into ten separate files, this allows up to 10 mini batches.

本配套示例脚本由加拿大安大略省汉密尔顿市麦克马斯特大学的卡洛斯·维德尔（Carlos Vidal）博士与菲利普·科利迈尔（Phillip Kollmeyer）博士开发。若将本脚本及配套数据集用于任何用途，需同时引用以下论文与该Mendeley数据集：

Carlos Vidal, Pawel Malysz, Mina Naguib, Ali Emadi, Phillip J. Kollmeyer, "Estimating battery state of charge using recurrent and non-recurrent neural networks," Journal of Energy Storage, 2021 (见https://www.sciencedirect.com/ 获取完整引用信息)。

本示例脚本用于训练两种不同类型的机器学习荷电状态（state of charge, SOC）估算算法：一类为带滤波输入值的前馈神经网络（feedforward neural network, FNN），另一类为长短期记忆（long short term memory, LSTM）循环神经网络。上述两种算法的详细说明已刊载于上述参考文献中。脚本针对归一化数据训练SOC估算器，所用数据集为松下18650PF动力电池数据集，可通过以下链接获取：https://data.mendeley.com/datasets/wykht8y7tg/1

%*****************************************************************************************%
-脚本下载与运行指南：
1. 从Mendeley Data页面下载全部文件
2. 下载的文件将以压缩包形式提供，请将其解压至指定文件夹，请勿修改文件夹结构
3. 打开并运行"LSTMvsFNN_Script_Nov_2021_V2.mlx"
4. 脚本内的注释与文本已包含进一步操作说明（如需获得更佳使用体验，请使用扩展名为*.mlx的Matlab实时脚本文件（Matlab Live Code File））。

%*****************************************************************************************%
重要提示：关于脚本的补充说明
如需切换神经网络类型，请前往第101行并按照以下指引操作。
选择神经网络类型（第101行）：
在下拉菜单中选择"1"以使用LSTM网络
在下拉菜单中选择"2"以使用FNN
同时请前往第237行，完成相同的神经网络类型选择。

%*****************************************************************************************%
配套文件说明：
- 所有数据均已完成归一化处理，如需了解反归一化方法，请参见"Normalization"文件夹中的示例说明
- 数据集包含-20、-10、0、10及25℃下的测试数据（本论文未使用-20℃数据集）
- "X"数据包含7行数据，各行数据依次为：{电压V、电流I、温度T、0.5mHz滤波电压V_0.5mHz、0.5mHz滤波电流I_0.5mHz、5mHz滤波电压V_5mHz、5mHz滤波电流I_5mHz}，其中V为电池电压，I为充放电电流，T为电池温度，带_0.5mHz与_5mHz后缀的数据为经过一阶低通巴特沃斯滤波器（Butterworth filter）处理的滤波值
- "Y"数据为通过库仑计数法（coulomb counting）计算得到的电池荷电状态
- 数据已划分为训练集、测试集与验证集，并分别存储于对应文件夹中（数据划分的具体规则详见《储能期刊》刊载的论文）。训练数据使用Mix 1至4与US06工况，验证数据使用LA92工况（-10℃），测试数据使用LA92、NN与UDDS工况
- 训练数据被拆分为10个独立文件，以支持最多10个小批量批次（mini batches）的训练